The vld1-lane, vld1-dup and vst1-lane instructions do not yet support using
post-increment versions, but all the rest of the NEON load/store instructions
should be handled now.
llvm-svn: 125014
These operations are expanded to pairs of loads or stores, and the first one
uses the address register update to produce the address for the second one.
So far, the second load/store has also updated the address register, just
for convenience, since that output has never been used. In anticipation of
actually supporting post-increment updates for these operations, this changes
the non-updating operations to use a non-updating load/store for the second
instruction.
llvm-svn: 125013
TargetInstrInfo:
Change produceSameValue() to take MachineRegisterInfo as an optional argument.
When in SSA form, targets can use it to make more aggressive equality analysis.
Machine LICM:
1. Eliminate isLoadFromConstantMemory, use MI.isInvariantLoad instead.
2. Fix a bug which prevent CSE of instructions which are not re-materializable.
3. Use improved form of produceSameValue.
ARM:
1. Teach ARM produceSameValue to look pass some PIC labels.
2. Look for operands from different loads of different constant pool entries
which have same values.
3. Re-implement PIC GA materialization using movw + movt. Combine the pair with
a "add pc" or "ldr [pc]" to form pseudo instructions. This makes it possible
to re-materialize the instruction, allow machine LICM to hoist the set of
instructions out of the loop and make it possible to CSE them. It's a bit
hacky, but it significantly improve code quality.
4. Some minor bug fixes as well.
With the fixes, using movw + movt to materialize GAs significantly outperform the
load from constantpool method. 186.crafty and 255.vortex improved > 20%, 254.gap
and 176.gcc ~10%.
llvm-svn: 123905
movw r0, :lower16:(L_foo$non_lazy_ptr-(LPC0_0+4))
movt r0, :upper16:(L_foo$non_lazy_ptr-(LPC0_0+4))
LPC0_0:
add r0, pc, r0
It's not yet enabled by default as some tests are failing. I suspect bugs in
down stream tools.
llvm-svn: 123619
instruction based on the t_addrmode_s# mode and what it returned. There is some
obvious badness to this. In particular, it's hard to do MC-encoding when the
instruction may change out from underneath you after the t_addrmode_s# variable
is finally resolved.
The solution is to revert a long-ago change that merged the reg/reg and reg/imm
versions. There is the addition of several new addressing modes. They no longer
have extraneous operands associated with them. I.e., if it's reg/reg we don't
have to have a dummy zero immediate tacked on to the SDNode.
There are some obvious cleanups here, which will happen shortly.
llvm-svn: 121747
Alignments smaller than the total size of the memory being loaded or stored,
unless the alignment is 8 bytes, are not allowed. Add tests for this, too.
llvm-svn: 121506
difficult on current ARM implementations for a few reasons.
1. Even though a single vmla has latency that is one cycle shorter than a pair
of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause
additional pipeline stall. So it's frequently better to single codegen
vmul + vadd.
2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to
stall for 4 cycles. We need to schedule them apart.
3. A vmla followed vmla is a special case. Obvious issuing back to back RAW
vmla + vmla is very bad. But this isn't ideal either:
vmul
vadd
vmla
Instead, we want to expand the second vmla:
vmla
vmul
vadd
Even with the 4 cycle vmul stall, the second sequence is still 2 cycles
faster.
Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough
but it isn't the optimial solution. This patch attempts to make it possible to
use vmla / vmls in cases where it is profitable.
A. Add missing isel predicates which cause vmla to be codegen'ed.
B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to
compute a fmul and a fmla.
C. Add additional isel checks for vmla, avoid cases where vmla is feeding into
fp instructions (except for the #3 exceptional case).
D. Add ARM hazard recognizer to model the vmla / vmls hazards.
E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the
vmla / vmls will trigger one of the special hazards.
Work in progress, only A+B are enabled.
llvm-svn: 120960
explicit about the operands. Split out the different variants into separate
instructions. This gives us the ability to, among other things, assign
different scheduling itineraries to the variants. rdar://8477752.
llvm-svn: 117409
which require the use of the shifter-operand. This will be used to split
the ldr/str instructions such that those versions needing the shifter operand
can get a different scheduling itenerary, as in some cases, the use of the
shifter can cause different scheduling than the simpler forms.
llvm-svn: 115066
passed the root of the match, even though only a few patterns
actually needed this (one in X86, several in ARM [which should
be refactored anyway], and some in CellSPU that I don't feel
like detangling). Instead of requiring all ComplexPatterns to
take the dead root, have targets opt into getting the root by
putting SDNPWantRoot on the ComplexPattern.
llvm-svn: 114471
instructions prior to regalloc. Since it's getting a little close to
the 2.8 branch deadline, I'll have to leave the rest of the instructions
handled by the NEONPreAllocPass for now, but I didn't want to leave half
of the VLD instructions converted and the other half not.
llvm-svn: 112983
int x(int t) {
if (t & 256)
return -26;
return 0;
}
We generate this:
tst.w r0, #256
mvn r0, #25
it eq
moveq r0, #0
while gcc generates this:
ands r0, r0, #256
it ne
mvnne r0, #25
bx lr
Scandalous really!
During ISel time, we can look for this particular pattern. One where we have a
"MOVCC" that uses the flag off of a CMPZ that itself is comparing an AND
instruction to 0. Something like this (greatly simplified):
%r0 = ISD::AND ...
ARMISD::CMPZ %r0, 0 @ sets [CPSR]
%r0 = ARMISD::MOVCC 0, -26 @ reads [CPSR]
All we have to do is convert the "ISD::AND" into an "ARM::ANDS" that sets [CPSR]
when it's zero. The zero value will all ready be in the %r0 register and we only
need to change it if the AND wasn't zero. Easy!
llvm-svn: 112664
all the other LDM/STM instructions. This fixes asm printer crashes when
compiling with -O0. I've changed one of the NEON tests (vst3.ll) to run
with -O0 to check this in the future.
Prior to this change VLDM/VSTM used addressing mode #5, but not really.
The offset field was used to hold a count of the number of registers being
loaded or stored, and the AM5 opcode field was expanded to specify the IA
or DB mode, instead of the standard ADD/SUB specifier. Much of the backend
was not aware of these special cases. The crashes occured when rewriting
a frameindex caused the AM5 offset field to be changed so that it did not
have a valid submode. I don't know exactly what changed to expose this now.
Maybe we've never done much with -O0 and NEON. Regardless, there's no longer
any reason to keep a count of the VLDM/VSTM registers, so we can use
addressing mode #4 and clean things up in a lot of places.
llvm-svn: 112322
with the VST4 instructions. Until after register allocation, we want to
represent sets of adjacent registers by a single super-register. These
VST4 pseudo instructions have a single QQ or QQQQ source register operand.
They get expanded to the real VST4 instructions with 4 separate D register
operands. Once this conversion is complete, we'll be able to remove the
NEONPreAllocPass and avoid some fragile and hacky code elsewhere.
llvm-svn: 112108
VECTOR_SHUFFLEs to REG_SEQUENCE instructions. The standard ISD::BUILD_VECTOR
node corresponds closely to REG_SEQUENCE but I couldn't use it here because
its operands do not get legalized. That is pretty awful, but I guess it
makes sense for other targets. Instead, I have added an ARM-specific version
of BUILD_VECTOR that will have its operands properly legalized.
This fixes the rest of Radar 7872877.
llvm-svn: 105439
A temporary flag -arm-tail-calls defaults to off,
so there is no functional change by default.
Intrepid users may try this; simple cases work
but there are bugs.
llvm-svn: 105413
const_casts, and it reinforces the design of the Target classes being
immutable.
SelectionDAGISel::IsLegalToFold is now a static member function, because
PIC16 uses it in an unconventional way. There is more room for API
cleanup here.
And PIC16's AsmPrinter no longer uses TargetLowering.
llvm-svn: 101635
These instructions are only needed for codegen, so I've removed all the
explicit encoding bits for now; they should be set in the same way as the for
VLDMD and VSTMD whenever we add encodings for VFP. The use of addrmode5
requires that the instructions be custom-selected so that the number of
registers can be set in the AM5Opc value.
llvm-svn: 99309
with changes to add a separate optional register update argument. Change all
the NEON instructions with address register writeback to use it.
llvm-svn: 99095
load/stores with address register writeback, and use "odd" suffix to distinguish
instructions to access odd numbered registers (instead of "a" and "b").
No functional changes.
llvm-svn: 99066
DoInstructionSelection. Inline "SelectRoot" into it from DAGISelHeader.
Sink some other stuff out of DAGISelHeader into SDISel.
Eliminate the various 'Indent' stuff from various targets, which dates
to when isel was recursive.
17 files changed, 114 insertions(+), 430 deletions(-)
llvm-svn: 97555
IsLegalToFold and IsProfitableToFold. The generic version of the later simply checks whether the folding candidate has a single use.
This allows the target isel routines more flexibility in deciding whether folding makes sense. The specific case we are interested in is folding constant pool loads with multiple uses.
llvm-svn: 96255
into TargetOpcodes.h. #include the new TargetOpcodes.h
into MachineInstr. Add new inline accessors (like isPHI())
to MachineInstr, and start using them throughout the
codebase.
llvm-svn: 95687
clear what information these functions are actually using.
This is also a micro-optimization, as passing a SDNode * around is
simpler than passing a { SDNode *, int } by value or reference.
llvm-svn: 92564
than doing the same via constpool:
1. Load from constpool costs 3 cycles on A9, movt/movw pair - just 2.
2. Load from constpool might stall up to 300 cycles due to cache miss.
3. Movt/movw does not use load/store unit.
4. Less constpool entries => better compiler performance.
This is only enabled on ELF systems, since darwin does not have needed
relocations (yet).
llvm-svn: 89720
naming scheme used in SelectionDAG, where there are multiple kinds
of "target" nodes, but "machine" nodes are nodes which represent
a MachineInstr.
llvm-svn: 82790
The instructions can be selected directly from the intrinsics. We will need
to add some ARM-specific nodes for VLD/VST of 3 and 4 128-bit vectors, but
those are not yet implemented.
llvm-svn: 80117
target-specific VDUPLANE nodes. This allows the subreg handling for the
quad-register version to be done easily with Pats in the .td file, instead
of with custom code in ARMISelDAGToDAG.cpp.
llvm-svn: 78993
match base only address, i.e. [r] since Thumb2 requires a offset register field.
For those, use [r + imm12] where the immediate is zero.
Note the generated assembly code does not look any different after the patch.
But the bug would have broken the JIT (if there is Thumb2 support) and it can
break later passes which expect the address mode to be well-formed.
llvm-svn: 78658
the overloaded vector types allowed floating-point or integer vector elements.
Most of these operations actually depend on the element type, so bitcasting
was not an option.
If you include the vpadd intrinsics that I updated earlier, this gets rid
of 20 intrinsics.
llvm-svn: 78646
directly from the intrinsics produced by the frontend. If it is more
convenient to have a custom DAG node for using these to implement shuffles,
we can add that later.
llvm-svn: 78459
This patch takes pain to ensure all the PEI lowering code does the right thing when lowering frame indices, insert code to manipulate stack pointers, etc. It's also custom lowering dynamic stack alloc into pseudo instructions so we can insert the right instructions at scheduling time.
This fixes PR4659 and PR4682.
llvm-svn: 78361
Get rid of yesterday's code to fix the register usage during isel.
Select the new DAG nodes to machine instructions. The new pre-alloc pass
to choose adjacent registers for these results is not done, so the
results of this will generally not assemble yet.
llvm-svn: 78136
After much back and forth, I decided to deviate from ARM design and split LDR into 4 instructions (r + imm12, r + imm8, r + r << imm12, constantpool). The advantage of this is 1) it follows the latest ARM technical manual, and 2) makes it easier to reduce the width of the instruction later. The down side is this creates more inconsistency between the two sub-targets. We should split ARM LDR instruction in a similar fashion later. I've added a README entry for this.
llvm-svn: 74420
predicate does not check if Thumb mode is enabled, and when in ARM mode
there are still some checks for constant-pool use that need to run.
llvm-svn: 73887
When compiling in Thumb mode, only the low (R0-R7) registers are available
for most instructions. Breaking the low registers into a new register class
handles this. Uses of R12, SP, etc, are handled explicitly where needed
with copies inserted to move results into low registers where the rest of
the code generator can deal with them.
llvm-svn: 68545
and into the ScheduleDAGInstrs class, so that they don't get
destructed and re-constructed for each block. This fixes a
compile-time hot spot in the post-pass scheduler.
To help facilitate this, tidy and do some minor reorganization
in the scheduler constructor functions.
llvm-svn: 62275
priority function. Instead, just iterate over the AllNodes list, which is
already in topological order. This eliminates a fair amount of bookkeeping,
and speeds up the isel phase by about 15% on many testcases.
The impact on most targets is that AddToISelQueue calls can be simply removed.
In the x86 target, there are two additional notable changes.
The rule-bending AND+SHIFT optimization in MatchAddress that creates new
pre-isel nodes during isel is now a little more verbose, but more robust.
Instead of either creating an invalid DAG or creating an invalid topological
sort, as it has historically done, it can now just insert the new nodes into
the node list at a position where they will be consistent with the topological
ordering.
Also, the address-matching code has logic that checked to see if a node was
"already selected". However, when a node is selected, it has all its uses
taken away via ReplaceAllUsesWith or equivalent, so it won't recieve any
further visits from MatchAddress. This code is now removed.
llvm-svn: 58748
flag. Then in a debugger developers can set breakpoints at these calls
to see waht is about to be selected and what the resulting subgraph
looks like. This really helps when debugging instruction selection.
llvm-svn: 58278
Instead, just create one, and make sure everything that needs
it can access it. Previously most of the SelectionDAGISel
subclasses all had their own TargetLowering object, which was
redundant with the TargetLowering object in the TargetMachine
subclasses, except on Sparc, where SparcTargetMachine
didn't have a TargetLowering object. Change Sparc to work
more like the other targets here.
llvm-svn: 57016
process up to a higher level. This allows FastISel to leverage
more of SelectionDAGISel's infastructure, such as updating Machine
PHI nodes.
Also, implement transitioning from SDISel back to FastISel in
the middle of a block, so it's now possible to go back and
forth. This allows FastISel to hand individual CallInsts and other
complicated things off to SDISel to handle, while handling the rest
of the block itself.
To help support this, reorganize the SelectionDAG class so that it
is allocated once and reused throughout a function, instead of
being completely reallocated for each block.
llvm-svn: 55219
replacement of multiple values. This is slightly more efficient
than doing multiple ReplaceAllUsesOfValueWith calls, and theoretically
could be optimized even further. However, an important property of this
new function is that it handles the case where the source value set and
destination value set overlap. This makes it feasible for isel to use
SelectNodeTo in many very common cases, which is advantageous because
SelectNodeTo avoids a temporary node and it doesn't require CSEMap
updates for users of values that don't change position.
Revamp MorphNodeTo, which is what does all the work of SelectNodeTo, to
handle operand lists more efficiently, and to correctly handle a number
of corner cases to which its new wider use exposes it.
This commit also includes a change to the encoding of post-isel opcodes
in SDNodes; now instead of being sandwiched between the target-independent
pre-isel opcodes and the target-dependent pre-isel opcodes, post-isel
opcodes are now represented as negative values. This makes it possible
to test if an opcode is pre-isel or post-isel without having to know
the size of the current target's post-isel instruction set.
These changes speed up llc overall by 3% and reduce memory usage by 10%
on the InstructionCombining.cpp testcase with -fast and -regalloc=local.
llvm-svn: 53728
and better control the abstraction. Rename the type
to MVT. To update out-of-tree patches, the main
thing to do is to rename MVT::ValueType to MVT, and
rewrite expressions like MVT::getSizeInBits(VT) in
the form VT.getSizeInBits(). Use VT.getSimpleVT()
to extract a MVT::SimpleValueType for use in switch
statements (you will get an assert failure if VT is
an extended value type - these shouldn't exist after
type legalization).
This results in a small speedup of codegen and no
new testsuite failures (x86-64 linux).
llvm-svn: 52044
Added ISD::DECLARE node type to represent llvm.dbg.declare intrinsic. Now the intrinsic calls are lowered into a SDNode and lives on through out the codegen passes.
For now, since all the debugging information recording is done at isel time, when a ISD::DECLARE node is selected, it has the side effect of also recording the variable. This is a short term solution that should be fixed in time.
llvm-svn: 46659
and StoreSDNode into their common base class LSBaseSDNode. Member
functions getLoadedVT and getStoredVT are replaced with the common
getMemoryVT to simplify code that will handle both loads and stores.
llvm-svn: 46538
that "machine" classes are used to represent the current state of
the code being compiled. Given this expanded name, we can start
moving other stuff into it. For now, move the UsedPhysRegs and
LiveIn/LoveOuts vectors from MachineFunction into it.
Update all the clients to match.
This also reduces some needless #includes, such as MachineModuleInfo
from MachineFunction.
llvm-svn: 45467
use ISD::{S,U}DIVREM and ISD::{S,U}MUL_HIO. Move the lowering code
associated with these operators into target-independent in LegalizeDAG.cpp
and TargetLowering.cpp.
llvm-svn: 42762